DeepSeek's logo can be seen on smartphone screens.
Omer Taha Cetin/Anadolu via Getty Images
DeepSeek, a Chatgpt competitor developed by Chinese hedge fund Mogul Liang Wenfeng, has climbed to the top of the download charts and started a debate over how a relatively obscure company has managed to build an advanced AI product.
Powered by a proprietary AI model, DeepSeek-Aim Assistant is now the most popular free app in the Apple App Store in China and the United States, according to San Francisco-based market intelligence platform Sensor Tower. Similar to Chatgpt, chatbots can perform tasks such as answering questions, drafting content, and helping users gather information.
The app's popularity has surged as DeepSeek, the Hangzhou company behind it, has troubled Silicon Valley with its technological breakthroughs. It appears to have developed advanced AI at a fraction of the cost that San Francisco-based Openai spent as U.S. officials imposed restrictions on the sale of semiconductors to China.
The DeepSeek R1 model released earlier this month is now the fourth most popular AI model in the world, according to Chatbot Arena, an evaluation platform developed by tech researchers including the University of California, Berkeley. According to Chatbot Arena, the product precedes the OpenAI O1 model released last December.
When it launched the product earlier this month, DeepSeek claimed it was capable of sophisticated reasoning and solving complex mathematical problems with performance that matched OpenAi O1. Liang met with Chinese Prime Minister Li Qiang last week as part of a government meeting to discuss AI.
The company did not respond to a request for further comment. The early success of DeepSeek, founded in 2023 with no big-name outside funding, has impressed Silicon Valley. Billionaire investor Marc Andreessen wrote in a January 24 post on the X social media platform: “The DeepSeek R1 is the most amazing and impressive thing I have ever seen. One of the profound breakthroughs of open source is a profound gift to the world.”
Paul Triolo, a partner at DGA-Albright Stonebridge Group in Washington, D.C., said by email that DeepSeek's model was considered the best, especially when benchmarking OpenAI's performance.
The company's meteoric rise has been driven by Liang Wenfeng, a reclusive hedge fund tycoon with a fixation on advanced technology. Liang's father was born in 1985 and is a primary school teacher. According to local media reports, he began his career as an investor. Liang began researching how to use AI to trade stocks in 2008, after earning a master's degree in Computer Vision at Hangzhou University.
Liang co-founded the hedge fund High-Flyer Quant in 2015 and began deploying its algorithms to evaluate investment opportunities in 2016. By 2017, nearly all High-Flyer Quant investments were made by AI, the company said on its website.
Two years later, High-Flyer Quant’s assets under management (AUM) reached 10 billion yuan ($1.4 billion). Some of its best-performing funds have returned more than 200% through 2024, according to data from China hithink flushing information cited by local media in China.
However, Liang's real interest seems to be in exploring cutting-edge technology. In a rare 2023 interview, the entrepreneur told local media 36KR that he wanted to research and innovate. While running high flight volumes, Liang began stockpiling NVIDIA chips before export controls began in Washington. He also invested his hedge fund returns into training AI models.
In 2023, DeepSeek was established. The startup remains self-funded and has no outside investors, according to business information service Qichacha. The company launched its first AI model in November 2023.
Xu Huazhe, assistant professor of information science at China's prestigious Tsinghua University, said via Wechat that some of DeepSeek's innovations appear to come from a hybrid technology called experts (MOE). It allows researchers to train many smaller AI models simultaneously and combine selected outputs in response to users.
Xu said the technology may not require many advanced chips or collect as much data, meaning DeepSeek could save costs. The company claims that some of these AI models only cost $5.6 million to train, which is reportedly 95% less than what OpenAI spent.
DeepSeek in turn charges users less. The AI Assistant Chatbot is free to use, but app developers pay to access the base model, which they can tweak to build other products based on. DeepSeek R1 charges 14 cents per million tokens, which in AI refers to the amount of data processed, while OpenAI charges $7.50.
Boutique Investment Bank Chanson & Co. Shen Meng, Beijing managing director of Via Wechat, said DeepSeek has charted its own technological path, while most Chinese models are imitators. He said this also aroused the hope of Chinese chip developers.
“Training of DeepSeek does not rely on much computing power, which means it does not require many of Nvidia's graphics processing units (GPUs, a type of AI chip),” Shen said. “This means that chips from China now have more room to grow.”
